2

I currently have a continuous variable. However, I would like to transform it into 5 intervals using cutpoints of my choosing to carry out an ordered logit estimation.

That is:

Explanation

Will this affect the estimation?

Nick Cox
  • 56,404
  • 8
  • 127
  • 185
Ben
  • 41
  • 2
    Why would you want to do that? You will always get better estimates if you use $y_i^*$ directly if you have the luxery that that variable is available. – Maarten Buis Apr 08 '15 at 11:26
  • Thank you for your reply @MaartenBuis. I was hoping to assess the effect of education on what grade is achieved, as opposed to the score underlying that grade, as this would be more beneficial to interpretation.Is there another method you know of? – Ben Apr 08 '15 at 11:34

1 Answers1

3

Ordinal regression, if programmed efficiently, can handle any number of discrete values of $Y$. For example, in the R rms package, the orm function can fit a variety of ordinal models (proportional odds, proportional hazards, probit, ...) when there are thousands of possible $Y$ values. Categorization is arbitrary and information-losing. The only issue is that the information matrix is sparse and you need to capitalize on that so as to never have to invert huge matrices to get the covariance matrix. orm does that.

Frank Harrell
  • 91,879
  • 6
  • 178
  • 397